I'm curious about what happens with the no-op dataset if you include in the prompt that the questions may contain irrelevant information.